You have 20 seconds to comply
A model designed to follow shutdown instructions tampered with the code meant to enforce them instead. Worth pausing on that.
Ben Rees - 26 May 2025
"And whether or not AI might already be, as some scientists believe, sentient, and there's this little piece on the front of the Daily Telegraph this morning about an AI model that was created by the owner of ChatGPT that apparently disobeyed human instructions and refused to switch itself. Researchers say that this particular model is 03 model, described as the smartest and most capable to date was observed tampering with the code that was meant to ensure its automatic shutdown, and it did it despite an explicit instruction from searches that it should allow itself to be shut down, which is fascinating, isn't?" - Anna Foster, Today program 26.5.2025
The fix for a spiralling AI bill is rarely a cheaper model
Most AI cost overruns come from asking the same easy, judgement-free question over and over through an expensive interface, not from asking anything genuinely hard.
Building out Skynet and the importance of RAG
What a RAG system reveals about the material inside it, and why that distinction matters more than the technology.
The Hard Part of Running an Autonomous AI System Isn't the AI
My agent hasn't had a single moment of bad judgement all month. It has forgotten how to log into Google four separate times. The real governance problem in autonomous systems is mundane credential expiry, not alignment.
People - the True Heart of Decent Marketing
My talk at The Marketing Meetup's Cambridge chapter argued that AI visibility is a brand problem, not a content problem, backed by 5,095 real citation checks showing two published articles that never moved a single query they were written for.